In temporal difference learning, positive prediction errors shift forward in time to the earliest reliable predictor of a primary reward